Tao Zhang, Assistant General Manager of the Bank for International Settlements, delivered a speech at the opening plenary of the International Financial Week in Hong Kong on the twenty-sixth of January in which he articulated, with unusual directness for a figure occupying a role at the apex of global financial governance, a set of concerns regarding the capacity of artificial intelligence and allied digital technologies to accelerate and amplify the transmission mechanisms through which financial stress propagates across institutional and geographical boundaries. Specifically, Zhang identified three distinct channels through which these technologies might operate to destabilise the financial system in a manner that proved more rapid or more intense than had historically been the case: the acceleration of trading and portfolio reallocation decisions; the concentration of critical operational functions across a limited number of technology providers and platforms; and the convergence of trading algorithms and risk management practices such that many institutions would respond to market stress in broadly similar ways, rather than in ways that might provide some degree of mutual stabilisation.

The BIS has become, in the years since the financial crisis of 2008 to 2009, the principal institutional source of comparative analysis regarding the architecture of the global financial system. Whilst the International Monetary Fund concerns itself principally with the surveillance of individual economies, and the Financial Stability Board with the harmonisation of regulatory standards across jurisdictions, the BIS has carved out an increasingly distinct role as the custodian of what might be termed the systemic properties of the financial order itself; that is, the characteristics and dynamics that determine the manner in which stress in one part of the system transmits to other parts, and the mechanisms through which feedback loops might amplify shocks rather than absorb them. Zhang's speech represented, in this regard, a formalisation within the official record of concerns that have circulated within regulatory and central banking circles for some months. Amongst the global institutions whose remit extends to financial stability, the BIS has been notably forward in raising questions about artificial intelligence; to do so with the directness evident in Zhang's remarks is to signal that the institution views the matter as warranting urgent attention from policymakers.

The first of the three channels identified concerns trading speed and the implications thereof. Contemporary financial markets operate at speeds that would be incomprehensible to market participants of any earlier epoch. The time required for a significant price movement in a major asset market to propagate through to price discovery in related markets has contracted from minutes and hours in the 1990s to milliseconds in the contemporary period. Artificial intelligence algorithms, inasmuch as they permit the identification of opportunities and the execution of responses at scales and speeds beyond human cognition, promise to accelerate this process still further. The implication, in Zhang's analysis, is that stress originating in one market or institution could transmit through interconnected markets with such rapidity that participants in the financial system would have insufficient time to absorb the shock, rebalance their positions, or implement other forms of risk mitigation; the result, in turn, could be the emergence of feedback loops in which selling pressure in one market generates additional selling pressure in others, with each successive round of liquidation occurring more rapidly than the preceding one, such that the total price movement could exceed what the underlying economic fundamentals would warrant.

The second risk channel concerns what has come to be known in the literature of financial regulation as operational concentration; that is, the condition in which critical functions of the financial system come to rest upon a small number of providers, such that a disruption affecting one provider could immediately cascade through large portions of the system. The rise of cloud computing platforms operated by a small number of major technology firms, and the migration of many financial institutions toward reliance upon these platforms for core payment, custody, and data processing functions, has created concentrations of operational risk that did not exist in earlier periods. Artificial intelligence systems, which are computationally intensive and require substantial investment in specialised infrastructure, have tended to reinforce this pattern of concentration; only a handful of entities globally command the resources required to operate large-scale machine learning systems, and so many financial institutions find themselves dependent upon these entities for access to the technology. The implication, as Zhang stated in his remarks, is that a technological failure affecting any one of a handful of major cloud providers or AI platforms could instantly disrupt the operations of hundreds or thousands of financial institutions simultaneously.

The third risk channel concerns what has come to be termed behavioural correlation; that is, the condition in which many financial institutions find themselves responding to the same information, or responding to information through the medium of broadly similar algorithms, in ways that generate correlated trading responses rather than idiosyncratic ones. In normal market conditions, the fact that different participants hold different views regarding asset values and future price movements provides a mechanism through which one participant's desire to sell can be offset against another's desire to buy, and the market can clear with a smaller adjustment in price than would be required if all participants held identical views and sought to take identical positions. However, when many institutions employ similar machine learning algorithms trained on similar data, they may converge upon similar conclusions regarding the direction of price movements and similar decisions regarding portfolio positioning, such that in periods of market stress the stabilising function of heterogeneity is lost and price movements are amplified.

The regulatory response to these three sets of risks is, at the present moment, still in its nascent phase. The Financial Stability Board has convened working groups to consider AI and financial stability, and individual regulatory authorities in major jurisdictions have begun the process of stress-testing financial institutions under scenarios in which AI-driven market dynamics produce shocks of varying magnitudes. However, no comprehensive international framework has yet emerged. The BIS's role, historically, has been to raise questions before the regulatory community has achieved consensus regarding the necessity for action; that pattern appears to be repeating in the context of artificial intelligence. The historical parallel most often cited by observers is the BIS's warnings regarding the financial stability implications of derivatives in the 1990s, issued at a time when the consensus view amongst central bankers and regulators was that derivatives markets, by allowing institutions to transfer risk, necessarily enhanced financial stability. That the BIS's concerns proved, in the fullness of time, to have been well founded is a matter of considerable historical irony; what remains to be determined is whether the institution's contemporaneous warnings regarding artificial intelligence will prove similarly prescient.